S^(hrdZddlmZddlZddlZddlmZddlmZm Z m Z m Z ddl Z ddlZddlmZddlmZmZmZmZmZmZmZmZmZdd lmZmZmZm Z m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)dd l*m+Z+m,Z,m-Z-dd l.m/Z/m0Z0m1Z1m2Z2m3Z3m4Z4d d l5m6Z6e3jne8Z9dZ:dZ;dZdZ?dZ@dZAdZBdZCdZDdZEdZFGddZGGdde'jjZJGdde'jjZKGd d!e'jjZLGd"d#e'jjZMGd$d%e'jjZNGd&d'e'jjZOGd(d)e'jjZPGd*d+e'jjZQGd,d-e'jjZRGd.d/e'jjZSGd0d1e'jjZTGd2d3e'jjZUGd4d5e'jjZVe(Gd6d7e'jjZWGd8d9e"ZXeGd:d;e/ZYdeZGd?d@eXZ\e1dAeZGdBdCeXeGZ]e1dDeZGdEdFeXeZ^GdGdHeXeZ_e1dIeZGdJdKeXe!Z`e1dLeZGdMdNeXe$Zae1dOeZGdPdQeXe Zbe1dReZGdSdTeXe%Zce1dUeZGdVdWeXe#ZdgdXZey)YzTF 2.0 BERT model.) annotationsN) dataclass)DictOptionalTupleUnion)get_tf_activation) +TFBaseModelOutputWithPastAndCrossAttentions.TFBaseModelOutputWithPoolingAndCrossAttentions#TFCausalLMOutputWithCrossAttentionsTFMaskedLMOutputTFMultipleChoiceModelOutputTFNextSentencePredictorOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFCausalLanguageModelingLossTFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFNextSentencePredictionLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax) ModelOutputadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) BertConfigzgoogle-bert/bert-base-uncasedr+z0dbmdz/bert-large-cased-finetuned-conll03-englishzZ['O', 'I-ORG', 'I-ORG', 'I-ORG', 'O', 'O', 'O', 'O', 'O', 'I-LOC', 'O', 'I-LOC', 'I-LOC'] g{Gz?zydshieh/bert-base-cased-squad2z'a nice puppet'gp= ף@z'ydshieh/bert-base-uncased-yelp-polarityz 'LABEL_1'ceZdZdZddZy)TFBertPreTrainingLossz Loss function suitable for BERT-like pretraining, that is, the task of pretraining a language model by combining NSP + MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation. ctjjdtjjj}|t j j|d|d}t j|ddk7|j}||z}t j|t j|z }|t j j|d|d }t j|ddk7|j} || z} t j| t j| z } t j|| zd S) NT) from_logits reductionlabelsr)y_truey_predidtypenext_sentence_labelr*)r*) rlossesSparseCategoricalCrossentropy ReductionNONEtfnnrelucastr7 reduce_sumreshape) selfr3logitsloss_fnunmasked_lm_losses lm_loss_maskmasked_lm_lossesreduced_masked_lm_lossunmasked_ns_loss ns_loss_maskmasked_ns_lossreduced_masked_ns_losss g/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/bert/modeling_tf_bert.pyhf_compute_lossz%TFBertPreTrainingLoss.hf_compute_losscs-,,<<Y^YeYeYoYoYtYt<u%BEEJJvh7G,HQWXYQZ[wwvh/47?Q?W?WX - <!#/?!@2==Q]C^!^#"%%**V|A\!\zz03II4PPN)r3 tf.TensorrDrQreturnrQ)__name__ __module__ __qualname____doc__rOrPrNr/r/\s  QrPr/cXeZdZdZdfd ZddZ d ddZxZS) TFBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c Vt|di|||_|j|_|j|_|j |_t jj|jd|_ t jj|j|_ y)N LayerNormepsilonnameraterW)super__init__config hidden_sizemax_position_embeddingsinitializer_rangerlayersLayerNormalizationlayer_norm_epsr[Dropouthidden_dropout_probdropoutrCrckwargs __class__s rNrbzTFBertEmbeddings.__init__{s "6" !--'-'E'E$!'!9!988AVAV]h8i||++1K1K+L rPctjd5|jd|jj|j gt |j|_dddtjd5|jd|jj|j gt |j|_ dddtjd5|jd|j|j gt |j|_ ddd|jryd|_ t|ddetj|jj 5|jj#dd|jj gdddyy#1swY[xYw#1swYxYw#1swYxYw#1swYyxYw) Nword_embeddingsweight)r^shape initializertoken_type_embeddings embeddingsposition_embeddingsTr[)r= name_scope add_weightrc vocab_sizerdrrfrrtype_vocab_sizerurerwbuiltgetattrr[r^buildrC input_shapes rNr~zTFBertEmbeddings.builds ]], - //{{--t/?/?@+D,B,BC*DK ]]2 3 )-!{{22D4D4DE+D,B,BC*9*D & ]]0 1 '+!33T5E5EF+D,B,BC(7(D $  ::  4d + 7t~~223 L$$dD$++2I2I%JK L L 81       L Ls2A F<A G *AG?3G!<G GG!G*c>| | td|At||jjt j |j |}t|dd}|t j|d}|2t jt j||d|zd }t j |j|}t j |j|} ||z| z} |j| } |j| | } | S) z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. Nz5Need to provide either `input_ids` or `input_embeds`.)paramsindicesrdimsvaluer*)startlimitaxisinputsrtraining) ValueErrorr!rcrzr=gatherrrr"fill expand_dimsrangerwrur[rl) rC input_ids position_idstoken_type_ids inputs_embedspast_key_values_lengthrrposition_embedstoken_type_embedsfinal_embeddingss rNcallzTFBertEmbeddings.calls   !6TU U  *9dkk6L6L MIIT[[)LM /4  !WW+Q?N  >>5[^Nd=delmL))4+C+C\ZIIT-G-GQ_`(?:=NN>>1A>B<*.,0.2-1 & && *& , & + & &  & rPrYc^eZdZdfd ZddZ d ddZd dZxZS) TFBertSelfAttentionc t|d i||j|jzdk7r&t d|jd|jd|j|_t |j|jz |_|j|j z|_tj|j |_ tjj|jt|jd|_tjj|jt|jd|_tjj|jt|jd|_tjj'|j( |_|j,|_||_y) NrzThe hidden size (z6) is not a multiple of the number of attention heads ()queryunitskernel_initializerr^keyrr_rW)rarbrdnum_attention_headsrintattention_head_size all_head_sizemathsqrtsqrt_att_head_sizerrgDenserrfrrrrjattention_probs_dropout_probrl is_decoderrcrms rNrbzTFBertSelfAttention.__init__s "6"    : : :a ?#F$6$6#78''-'A'A&B!E  $*#=#= #&v'9'9F ."<&q)I(+K3N 11$((BW(2XZdeI33DJJF[J4\^hiK3N  '11$((-(2PR\]I33DJJmJ4TV`aK >!#4i"@qII))^A%6 $D1MK11$((-(2PR\]I33DJJmJ4TV`aK//0A:N ??(5N99[)N WWT,,4D4J4J K99%5r:  %!vv&6G )0@rJ,,o,Q   kk/9EO99_kB<<(8|L::-=jRTVZVhVhEij9J#_5QaPc ?? 11GrPc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swYxYw#1swYxYw#1swYyxYw)NTrrr) r|r}r=rxrr^r~rcrdrrrs rNr~zTFBertSelfAttention.build@s9 ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4 % 1txx}}- FdDKK,C,CDE F 4$ ' 3tzz/ H   $dkk.E.E!FG H H 4  H H F F H Hs$ 3E*<3E6-3F*E36E?F r)rrQrrrRrQF)rrQrrQrrQrrQrrQrTuple[tf.Tensor]rrrrrRrr)rSrTrUrbrrr~rrs@rNrrs87 O O"O O ) O !* O)O OO Ob HrPrc2eZdZdfd ZdddZddZxZS)TFBertSelfOutputc xt|di|tjj |j t |jd|_tjj|jd|_ tjj|j|_||_yNdenserr[r\r_rWrarbrrgrrdrrfrrhrir[rjrkrlrcrms rNrbzTFBertSelfOutput.__init__P "6"\\''$$IaIa9bip(  88AVAV]h8i||++1K1K+L  rPcz|j|}|j||}|j||z}|SNrrrrlr[rCr input_tensorrs rNrzTFBertSelfOutput.callZ? - 8  MH M ml.JK rPc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTrr[ r|r}r=rxrr^r~rcrdr[rs rNr~zTFBertSelfOutput.builda ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 L$$dD$++2I2I%JK L L 8 H H L L 3C9<3D9DDrrrrQrrQrrrRrQrrSrTrUrbrr~rrs@rNrrO LrPrc\eZdZdfd ZdZ d ddZddZxZS) TFBertAttentionc lt|di|t|d|_t |d|_y)NrCr^outputrW)rarbrself_attentionr dense_outputrms rNrbzTFBertAttention.__init__ns1 "6"1&vF,V(CrPctrNotImplementedError)rCheadss rN prune_headszTFBertAttention.prune_headsts!!rPc x|j||||||||} |j| d||} | f| ddz} | S)Nrrrrrrrrrrrrr*)rr) rCrrrrrrrr self_outputsrrs rNrzTFBertAttention.callwso**&)"7#9)/+   ,,&q/ x- $% QR(88rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)r|r}r=rxrr^r~rrs rNr~zTFBertAttention.builds ::  4)4 0 <t22778 0##))$/ 0 4 . :t00556 .!!''- . . ; 0 0 . . C %C CC rr)rrQrrQrrQrrQrrQrrrrrrrRrr)rSrTrUrbrrr~rrs@rNrrmsuD ""  )  !* )  : .rPrc0eZdZdfd ZddZddZxZS)TFBertIntermediatec Tt|di|tjj |j t |jd|_t|jtr"t|j|_ ||_y|j|_ ||_y)NrrrW)rarbrrgrintermediate_sizerrfr isinstance hidden_actstrr intermediate_act_fnrcrms rNrbzTFBertIntermediate.__init__s "6"\\''**vOgOg?hov(   f'' -'89J9J'KD $ (.'8'8D $ rPcL|j|}|j|}|SNr)rr rCrs rNrzTFBertIntermediate.calls( - 8 00? rPc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYwNTr r|r}r=rxrr^r~rcrdrs rNr~zTFBertIntermediate.build} ::  4$ ' 3tzz/ H   $dkk.E.E!FG H H 4 H H 3BBrrrQrRrQrrrs@rNrrs  HrPrc2eZdZdfd ZdddZddZxZS) TFBertOutputc xt|di|tjj |j t |jd|_tjj|jd|_ tjj|j|_||_yrrrms rNrbzTFBertOutput.__init__rrPcz|j|}|j||}|j||z}|Srrrs rNrzTFBertOutput.callrrPc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwr) r|r}r=rxrr^r~rcrr[rdrs rNr~zTFBertOutput.builds ::  4$ ' 3tzz/ N   $dkk.K.K!LM N 4d + 7t~~223 L$$dD$++2I2I%JK L L 8 N N L Lrrrrrrrs@rNrrrrPrcVeZdZdfd Z d ddZddZxZS) TFBertLayerc Dt|di|t|d|_|j|_|j |_|j r,|jst |dt|d|_t|d|_ t|d|_ y)N attentionrz> should be used as a decoder model if cross attention is addedcrossattention intermediaterrW) rarbrrradd_cross_attentionrrrrr bert_outputrms rNrbzTFBertLayer.__init__s "6"(kB ++#)#=#=  # #?? D6)g!hii"1&?O"PD .vNK'X>rPc ||ddnd} |j|||dd| ||} | d} |jr | dd} | d} n| dd} d}|jrV|Tt|dstd|d||d dnd}|j | |||||||}|d} | |ddz} |d} |z} |j | }|j || | }|f| z} |jr|  fz} | S) Nr)rrrrrrrrrr*rrz'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`rr)rrhasattrrrrr )rCrrrrrrrrself_attn_past_key_valueself_attention_outputsrrpresent_key_valuecross_attn_present_key_valuecross_attn_past_key_valuecross_attention_outputsintermediate_output layer_outputs rNrzTFBertLayer.calls:H9S>"1#5Y] !%&)"&#'3/"0 " 2!4 ??,Qr2G 6r : ,QR0G'+$ ??4@4!12 =dVDDD @N?Yrs(;_c %&*&9&9--#&;'=8"3!': ' # 7q9  7" ==G,C2+F ( 14P P "//>N/O''-h hsAc | rdnd} |rdnd} |r|jjrdnd}|rdnd}t|jD]h\}}| r| |fz} |||nd}|||||||||| }|d}|r ||dfz }|s=| |dfz} |jjs]|`||dfz}j| r| |fz} | st d|| | |fDSt ||| | |S) NrWrrrr*rc3&K|] }|| ywrrW).0vs rN z%TFBertEncoder.call..zsghgts)last_hidden_statepast_key_valuesr attentionscross_attentions)rcr enumerater3tupler )rCrrrrrr; use_cacheroutput_hidden_states return_dictrall_hidden_statesall_attentionsall_cross_attentionsnext_decoder_cacher4 layer_moduler layer_outputss rNrzTFBertEncoder.callHsV#7BD0d%64;;;Z;Zr`d#,R$(4 VOA|#$58H$H!3B3N_Q/TXN(+-#A,&;'=-"3! M*!,M"}R'8&::" !/=3C2E!E;;227L7X+?=QRCSBU+U(1 V6 1]4D D )+bias_addrb)rCr seq_lengths rNrzTFBertLMPredictionHead.calls]C  .q1  -DDTDT?UV  MT5J5J5Q5Q_cd  -JPTP[P[PfPf?gh ]K rPrcr+r^keras.layers.LayerrrRrurz tf.Variable)rRzDict[str, tf.Variable]r) rSrTrUrbr~rfrjrlrnrrrs@rNr[r[s' 1+%@#>rPr[c0eZdZdfd ZddZddZxZS) TFBertMLMHeadc Jt|di|t||d|_y)N predictionsrrW)rarbr[r{r_s rNrbzTFBertMLMHead.__init__s& "6"1&:JQ^_rPc*|j|}|S)Nr#)r{)rCsequence_outputprediction_scoress rNrzTFBertMLMHead.calls ,,?,K  rPc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr{)r|r}r=rxr{r^r~rs rNr~zTFBertMLMHead.build sm ::  4 - 9t//445 -  &&t, - - : - - A11A:rt)r}rQrRrQrrrs@rNryrys` ! -rPryc0eZdZdfd ZddZddZxZS) TFBertNSPHeadc t|di|tjj dt |j d|_||_y)Nrseq_relationshiprrW) rarbrrgrrrfrrcrms rNrbzTFBertNSPHead.__init__sL "6" % 2 2.v/G/GH#!3!   rPc*|j|}|Sr )r)rCrRseq_relationship_scores rNrzTFBertNSPHead.call!s!%!6!6m!6!L%%rPc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYw)NTr) r|r}r=rxrr^r~rcrdrs rNr~zTFBertNSPHead.build&s ::  4+T 2 >t4499: S%%++T49P9P,QR S S ? S Srr)rRrQrRrQrrrs@rNrrs& SrPrceZdZeZddfd Zd dZd dZdZe d d dZ d dZ xZ S)TFBertMainLayerc t|di|||_|j|_t |d|_t |d|_|rt|d|_ yd|_ y)NrvrencoderpoolerrW) rarbrcrrYrvr0rrLr)rCrcadd_pooling_layerrnros rNrbzTFBertMainLayer.__init__3sZ "6"  ++*6 E$V)< =Nl69 TX rPc|jSr)rvres rNget_input_embeddingsz$TFBertMainLayer.get_input_embeddings=s rPc`||j_t|d|j_yrh)rvrrr"rzris rNset_input_embeddingsz$TFBertMainLayer.set_input_embeddings@s$!&%/%6q%9"rPct)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel r)rCheads_to_prunes rN _prune_headszTFBertMainLayer._prune_headsDs "!rPc |jjsd} | | td| t|}n|t|dd}n td|\}}| &d}dgt |j j z} nt| ddd}|tj|||zfd}|tj|d}|j|||||| }t|}||z}|jrtj|}tjtj|ddddf||df|ddddf}tj||j }||dddddfz}t|}tj||dd|d|d f}| d3|dddd| dddf}n tj||ddd|df}tj||j }tj d |j }tj d |j }tj"tj$|||}|jrf|dtj||j }t t|}|dk(r|dddddddf}|d k(r |ddddddf}d z d z}nd}|t&dg|jj(z}|j |||||| | | | | | }|d}|j*|j+|nd}| s ||f|ddzSt-|||j.|j0|j2|j4S)NFzDYou cannot specify both input_ids and inputs_embeds at the same timerz5You have to specify either input_ids or inputs_embedsrr"r*r)rrrrrrr6rg?gr ) rrrrrr;r@rrArBrr#)r: pooler_outputr;rr<r=)rcrrr"lenrr3r=rrvr less_equaltiler@r7rBconstantrsubtractrr2rr r;rr<r=) rCrrrrrrrrr;r@rrArBrrrrsrembedding_outputattention_mask_shapemask_seq_lengthseq_ids causal_maskextended_attention_maskone_cstten_thousand_cstnum_dims_encoder_attention_maskencoder_extended_attention_maskencoder_outputsr}rRs rNrzTFBertMainLayer.callKs${{%%I  ]%>cd d  "$Y/K  &$]3CR8KTU U!, J  "%& "#fs4<<+=+='>>O%/0B10E%Fr%J "  !WW:zDZ7Z*[cdeN  !WW+Q?N??%)'#9 +  *.9$'== ??hh/G--dA ._a0PQa &K''+^5I5IJK&1N1dA:4N&N ##-.E#F &(jj'*>q*A1FZ[\F]_stu_v)w' #q!-*A!Q VWBW*X'&(jj!5a!8!Q@TUV@W X' ##%''*AIYI_I_"`++c)9)?)?@;;x7G7M7MN"$++bkk'CZ.[]m"n ??5A&(WW-CKbKhKh%i ".1*=S2T.U +.!32HDRSUV2W/.!32HDRVXYIY2Z/035T/TX`.` +.2 +  % %!>!>>I,,*2"7#B+/!5#'  *!,FJkkF] / Bcg  #$ $ >-'+;;)77&11,==   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrvrr) r|r}r=rxrvr^r~rrrs rNr~zTFBertMainLayer.builds ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) 44 ( 4t{{//0 ( !!$' ( ( 5  , , ) ) ( ($ D%%D1?D=%D.1D:=ETrcr+rrrvrwNNNNNNNNNNNNNF)rTFModelInputType | Nonernp.ndarray | tf.Tensor | Nonerrrrrrrrrrrrr;4Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]]r@rJrrJrArJrBrJrrrRGUnion[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]r) rSrTrUr+ config_classrbrrrr rr~rrs@rNrr/s LY:".28<8<6:377;?C@DPT$(,0/3&*[ *[ 6[ 6 [ 4 [ 1 [ 5[  =[ !>[ N[ "[ *[ -[ $[ [ Q![ [ z (rPrceZdZdZeZdZy)TFBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. bertN)rSrTrUrVr+rbase_model_prefixrWrPrNrrs LrPrcXeZdZUdZdZded<dZded<dZded<dZded <dZ ded <y) TFBertForPreTrainingOutputa? Output type of [`TFBertForPreTraining`]. Args: prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). seq_relationship_logits (`tf.Tensor` of shape `(batch_size, 2)`): Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax). hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Nr.lossrprediction_logitsseq_relationship_logitsz,Optional[Union[Tuple[tf.Tensor], tf.Tensor]]rr<) rSrTrUrVr__annotations__rrrr<rWrPrNrrsB,"D !-1*13707BFM?F?CJ<CrPrav This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior. TensorFlow models and layers in `transformers` accept two formats as input: - having all inputs as keyword arguments (like PyTorch models), or - having all inputs as a list, tuple or dict in the first positional argument. The reason the second format is supported is that Keras methods prefer this format when passing inputs to models and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first positional argument: - a single Tensor with `input_ids` only and nothing else: `model(input_ids)` - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])` - a dictionary with one or several input Tensors associated to the input names given in the docstring: `model({"input_ids": input_ids, "token_type_ids": token_type_ids})` Note that when creating models and layers with [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python function! Args: config ([`BertConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the model weights. a Args: input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) token_type_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) position_ids (`np.ndarray` or `tf.Tensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) head_mask (`np.ndarray` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`np.ndarray` or `tf.Tensor` of shape `({0}, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. training (`bool`, *optional*, defaults to `False``): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). z^The bare Bert Model transformer outputting raw hidden-states without any specific head on top.ceZdZddfd Zeeejdee e e  d d dZ d dZ xZS) TFBertModelcRt||g|i|t||d|_y)Nrr)rarbrr)rCrcrrrnros rNrbzTFBertModel.__init__s+ 3&3F3#F,=FK rPbatch_size, sequence_length checkpoint output_typercD|j||||||||| | | | | |}|S)a encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`) contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. use_cache (`bool`, *optional*, defaults to `True`): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). Set to `False` during training, `True` during generation rrrrrrrrr;r@rrArBr)r)rCrrrrrrrrr;r@rrArBrrs rNrzTFBertModel.callsIX))))%'"7#9+/!5#  rPc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr)r|r}r=rxrr^r~rs rNr~zTFBertModel.buildse ::  4 & 2tyy~~. & % & & 3 & &rrrr)rrrrrrrrrrrrrrrrr;rr@rJrrJrArJrBrJrrJrRrr)rSrTrUrbr r'BERT_INPUTS_DOCSTRINGformatr%_CHECKPOINT_FOR_DOCr _CONFIG_FOR_DOCrr~rrs@rNrrs L *+@+G+GHe+fg&B$.28<8<6:377;?C@DPT$(,0/3&*#(5*5656 5 4 5 1 555 =5!>5N5"5*5-5$5!5 Q!5 h5n&rPrz Bert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next sentence prediction (classification)` head. ceZdZgdZd fd Zd dZd dZeee jde e e  d d dZddZxZS)TFBertForPreTraining)rcls.predictions.decoder.weightzcls.predictions.decoder.biasct||g|i|t|d|_t |d|_t ||jjd|_y)Nrr nsp___cls mlm___clsr^r^) rarbrrrnspryrvmlmrCrcrrnros rNrbzTFBertForPreTraining.__init__sQ 3&3F3#F8  k: $)):N:NU`arPc.|jjSrrr{res rN get_lm_headz TFBertForPreTraining.get_lm_headxx###rPctjdt|jdz|jjzdz|jj jzSNzMThe method get_prefix_bias_name is deprecated. Please use `get_bias` instead./warningswarn FutureWarningr^rr{res rNget_prefix_bias_namez)TFBertForPreTraining.get_prefix_bias_nameG egtuyy3.4txx7K7K7P7PPPrPrrrc N|j||||||||| |  } | dd\}}|j|| }|j|}d}| | d| i}| |d<|j|||f}| s||f| ddz}||f|zS|St |||| j | j S) a labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` next_sentence_label (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see `input_ids` docstring) Indices should be in `[0, 1]`: - 0 indicates sequence B is a continuation of sequence A, - 1 indicates sequence B is a random sequence. kwargs (`Dict[str, any]`, *optional*, defaults to `{}`): Used to hide legacy arguments that have been deprecated. Return: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFBertForPreTraining >>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased") >>> model = TFBertForPreTraining.from_pretrained("google-bert/bert-base-uncased") >>> input_ids = tokenizer("Hello, my dog is cute", add_special_tokens=True, return_tensors="tf") >>> # Batch size 1 >>> outputs = model(input_ids) >>> prediction_logits, seq_relationship_logits = outputs[:2] ``` rrrrrrrrArBrNrr}rrRr3r8r3rD)rrrrr<)rrrrOrrr<)rCrrrrrrrrArBr3r8rrr}rRr~r total_lossd_labelsrs rNrzTFBertForPreTraining.calls`))))%'/!5#  *1!& HH_xHX!% !F  "5"A &)H.AH* +--XGXZpFq-rJ')?@712;NF/9/EZMF* Q6 Q)/$:!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrrr) r|r}r=rxrr^r~rrrs rNr~zTFBertForPreTraining.buildCs ::  4 & 2tyy~~. & % & 4 % 1txx}}- %t$ % 4 % 1txx}}- %t$ % % 2  & & % % % %rrrvrRr  NNNNNNNNNNNF)rrrrrrrrrrrrrrJrArJrBrJr3rr8rrrJrRz3Union[TFBertForPreTrainingOutput, Tuple[tf.Tensor]]r)rSrTrU"_keys_to_ignore_on_load_unexpectedrbrrr r'rrr)rrrr~rrs@rNrrs*& b$Q*+@+G+GHe+fg+ETcd.28<8<6:377;,0/3&*04=A#(M *M 6M 6 M 4 M 1 M 5M *M -M $M .M ;M !M  =M ehM ^ %rPrz2Bert Model with a `language modeling` head on top.c eZdZgdZd fd Zd dZd dZeee jde e e edd d dd Zdd ZxZS)TFBertForMaskedLMrcls.seq_relationshiprrct||g|i||jrtj dt |dd|_t||j jd|_ y)NzmIf you want to use `TFBertForMaskedLM` make sure `config.is_decoder=False` for bi-directional self-attention.Frrr^rr rarbrloggerwarningrrryrvrrs rNrbzTFBertForMaskedLM.__init__\sa 3&3F3    NN1  $Fe&Q  $)):N:NU`arPc.|jjSrrres rNrzTFBertForMaskedLM.get_lm_headhrrPctjdt|jdz|jjzdz|jj jzSrrres rNrz&TFBertForMaskedLM.get_prefix_bias_namekrrPrz'paris'g)\(?rrrexpected_output expected_lossc |j||||||||| |  } | d} |j| | }| dn|j| |}| s|f| ddz}||f|zS|St||| j| j S)a labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` rrrNrrrrDrr<)rrrOrrr<)rCrrrrrrrrArBr3rrr}r~rrs rNrzTFBertForMaskedLM.callos:))))%'/!5#  "!* HH_xHX~t4+?+?vVg+?+h')GABK7F)-)9TGf$ Ev E$!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwNTrrr|r}r=rxrr^r~rrs rNr~zTFBertForMaskedLM.build ::  4 & 2tyy~~. & % & 4 % 1txx}}- %t$ % % 2 & & % %rrrvr NNNNNNNNNNF)rrrrrrrrrrrrrrJrArJrBrJr3rrrJrRz)Union[TFMaskedLMOutput, Tuple[tf.Tensor]]r)rSrTrUrrbrrr r'rrr%rrrrr~rrs@rNrrRs*& b$Q*+@+G+GHe+fg&$$! .28<8<6:377;,0/3&*04#(- *- 6- 6 - 4 - 1 - 5- *- -- $- .- !-  3- h- ^ %rPrceZdZgdZd fd Zd dZd dZd dZee e e e  d ddZ ddZxZS)TFBertLMHeadModelrct||g|i||jstj dt |dd|_t||j jd|_ y)NzNIf you want to use `TFBertLMHeadModel` as a standalone, add `is_decoder=True.`Frrrrrrs rNrbzTFBertLMHeadModel.__init__s[ 3&3F3  NNk l#Fe&Q  $)):N:NU`arPc.|jjSrrres rNrzTFBertLMHeadModel.get_lm_headrrPctjdt|jdz|jjzdz|jj jzSrrres rNrz&TFBertLMHeadModel.get_prefix_bias_namerrPc n|j}|tj|}| |ddddf}|||dS)Nr)rrr;)rsr=ones)rCrr;r model_kwargsrs rNprepare_inputs_for_generationz/TFBertLMHeadModel.prepare_inputs_for_generationsEoo  !WW[1N  &!!RS&)I&.]lmmrPrc f|j||||||||| | | | | |}|d}|j||}d}|)|ddddf}|ddddf}|j||}| s|f|ddz}||f|zS|St|||j|j |j |j S) a encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`) contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. use_cache (`bool`, *optional*, defaults to `True`): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). Set to `False` during training, `True` during generation labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the cross entropy classification loss. Indices should be in `[0, ..., config.vocab_size - 1]`. rrrNrr*rr)rrDr;rr<r=)rrrOr r;rr<r=)rCrrrrrrrrr;r@rrArBr3rrnrr}rDrshifted_logitsrs rNrzTFBertLMHeadModel.calls`))))%'"7#9+/!5#  "!*/HM  #AssF^NAqrE]F''vn'MDY,F)-)9TGf$ Ev E2#33!//))$55   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwrrrs rNr~zTFBertLMHeadModel.build/rrrrvr)NN)NNNNNNNNNNNNNNF) rrrrrrrrrrrrrrrrr;rr@rJrrJrArJrBrJr3rrrJrRzO NO "O *O -O $O .O !!O $ F%O  O b %rPrzJBert Model with a `next sentence prediction (classification)` head on top.ceZdZddgZdfd Zeeejde e e  d d dZ d dZ xZS) TFBertForNextSentencePredictionrcls.predictionsctt||g|i|t|d|_t |d|_y)Nrrr)rarbrrrrrs rNrbz(TFBertForNextSentencePrediction.__init__Cs6 3&3F3#F8  k:rPrrc |j||||||||| |  } | d} |j| }| dn|j| |}| s|f| ddz}||f|zS|St||| j| j S)aB Return: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFBertForNextSentencePrediction >>> tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased") >>> model = TFBertForNextSentencePrediction.from_pretrained("google-bert/bert-base-uncased") >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced." >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light." >>> encoding = tokenizer(prompt, next_sentence, return_tensors="tf") >>> logits = model(encoding["input_ids"], token_type_ids=encoding["token_type_ids"])[0] >>> assert logits[0][0] < logits[0][1] # the next sentence was random ```rr*rNrrr)rrrOrrr<)rCrrrrrrrrArBr8rrrRseq_relationship_scoresnext_sentence_lossrs rNrz$TFBertForNextSentencePrediction.callIsH))))%'/!5#    "&((("G#* %%-@I`%a  -/'!"+=F7I7U')F2 a[a a,#*!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)r|r}r=rxrr^r~rrs rNr~z%TFBertForNextSentencePrediction.buildrrrr)rrrrrrrrrrrrrrJrArJrBrJr8rrrJrRz6Union[TFNextSentencePredictorOutput, Tuple[tf.Tensor]]r)rSrTrUrrbr r'rrr)rrrr~rrs@rNrr;s +78J)K&; *+@+G+GHe+fg+HWfg.28<8<6:377;,0/3&*=A#(> *> 6> 6 > 4 > 1 > 5> *> -> $> ;> !>  @> hh> @ %rPrz Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. c eZdZgdZdgZdfd Zeeejde e e e ee d d dZd dZxZS) TFBertForSequenceClassificationrrrrrlct||g|i||j|_t|d|_|j |j n |j }tjj||_ tjj|jt|jd|_||_y)Nrrr_ classifierrrarb num_labelsrrclassifier_dropoutrkrrgrjrlrrrfrrcrCrcrrnrros rNrbz(TFBertForSequenceClassification.__init__s 3&3F3 ++#F8 )/)B)B)NF % %TZTnTn ||++1C+D ,,,,##.v/G/GH-   rPrrc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) a labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rr*rrNrrr)rrlrrOrrr<)rCrrrrrrrrArBr3rrrRrDrrs rNrz$TFBertForSequenceClassification.calls:))))%'/!5#     MH M  6~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E)!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTrr r|r}r=rxrr^r~rrcrdrs rNr~z%TFBertForSequenceClassification.build ::  4 & 2tyy~~. & % & 4t , 8t334 M%%tT4;;3J3J&KL M M 9 & & M M C"%3C."C+.C7rr)rrrrrrrrrrrrrrJrArJrBrJr3rrrJrRz3Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]r)rSrTrUr_keys_to_ignore_on_load_missingrbr r'rrr%'_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATIONrr_SEQ_CLASS_EXPECTED_OUTPUT_SEQ_CLASS_EXPECTED_LOSSrr~rrs@rNrrs*s&'1l#"*+@+G+GHe+fg:.$2. .28<8<6:377;,0/3&*04#(. *. 6. 6 . 4 . 1 . 5. *. -. $. .. !.  =. h. ` MrPrz Bert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. ceZdZgdZdgZdfd Zeeejde e e e  d d dZd dZxZS) TFBertForMultipleChoicerrlc0t||g|i|t|d|_tj j |j|_tj jdt|jd|_ ||_ y)Nrrr_r*rr)rarbrrrrgrjrkrlrrrfrrcrs rNrbz TFBertForMultipleChoice.__init__s~ 3&3F3#F8 ||++1K1K+L ,,,,8P8P(QXd-  rPz(batch_size, num_choices, sequence_lengthrc |t|d} t|d} nt|d} t|d} |tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|&tj|d| t|dfnd}|j||||||||| |  }|d}|j || }|j | }tj|d| f}| dn|j | | }| s|f|ddz}||f|zS|St|||j|j S) aE labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) Nr*rrrr rrrrr) r"r=rBrrlrrOrrr<)rCrrrrrrrrArBr3r num_choicesrsflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrrRrDreshaped_logitsrrs rNrzTFBertForMultipleChoice.calls4  $Y/2K#I.q1J$]3A6K#M215JQZQf9RD uD L MrPr)z Bert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. c eZdZgdZdgZdfd Zeeejde e e e ee d d dZd dZxZS) TFBertForTokenClassificationrrrrrrlct||g|i||j|_t|dd|_|j |j n |j }tjj||_ tjj|jt|jd|_||_y)NFrrr_rrrrs rNrbz%TFBertForTokenClassification.__init__{s 3&3F3 ++#Fe&Q )/)B)B)NF % %TZTnTn ||++1C+D ,,,,##.v/G/GH-   rPrrc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) z labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. rrrrNrrr)rrlrrOrrr<)rCrrrrrrrrArBr3rrr}rDrrs rNrz!TFBertForTokenClassification.calls6))))%'/!5#  "!*,,o,Q8~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E&!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwr r!rs rNr~z"TFBertForTokenClassification.buildr"r#rr)rrrrrrrrrrrrrrJrArJrBrJr3rrrJrRz0Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]r)rSrTrUrr$rbr r'rrr%$_CHECKPOINT_FOR_TOKEN_CLASSIFICATIONrr_TOKEN_CLASS_EXPECTED_OUTPUT_TOKEN_CLASS_EXPECTED_LOSSrr~rrs@rNr5r5is*&(2l#"*+@+G+GHe+fg7+$40 .28<8<6:377;,0/3&*04#(, *, 6, 6 , 4 , 1 , 5, *, -, $, ., !,  :, h, \ MrPr5z Bert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). c eZdZgdZdfd Zeeejde e e e e eee d d dZd dZxZS) TFBertForQuestionAnsweringr6c t||g|i||j|_t|dd|_t j j|jt|jd|_ ||_ y)NFrr qa_outputsr) rarbrrrrrgrrrfr@rcrs rNrbz#TFBertForQuestionAnswering.__init__su 3&3F3 ++#Fe&Q ,,,,##.v/G/GH-   rPr)rrrqa_target_start_indexqa_target_end_indexrrc |j||||||||| |  } | d}|j|}tj|dd\}}tj|d}tj|d}d}| | d | i}| |d <|j |||f }| s||f| ddz}||f|zS|St |||| j| j S) a start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. rrrrr)rnum_or_size_splitsr)inputrNstart_position end_positionr)r start_logits end_logitsrr<) rr@r=splitsqueezerOrrr<)rCrrrrrrrrArBstart_positions end_positionsrrr}rDrHrIrr3rs rNrzTFBertForQuestionAnswering.callsH))))%'/!5#  "!*8#%88&QUW#X jzz 2> ZZjr:   &=+D&8F%2F> "''v|Z>X'YD"J/'!"+=F)-)9TGf$ Ev E-%!!//))   rPc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr@) r|r}r=rxrr^r~r@rcrdrs rNr~z TFBertForQuestionAnswering.build5r"r#rr)rrrrrrrrrrrrrrJrArJrBrJrLrrMrrrJrRz7Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]r)rSrTrUrrbr r'rrr%_CHECKPOINT_FOR_QArr_QA_TARGET_START_INDEX_QA_TARGET_END_INDEX_QA_EXPECTED_OUTPUT_QA_EXPECTED_LOSSrr~rrs@rNr>r>s*& *+@+G+GHe+fg%2$40+'.28<8<6:377;,0/3&*9=7;#(; *; 6; 6 ; 4 ; 1 ; 5; *; -; $; 7; 5; !;  A; h; z MrPr>) rYrr)rrr>rr5rrrr)frV __future__rrr dataclassesrtypingrrrrnumpynp tensorflowr=activations_tfr modeling_tf_outputsr r r rrrrrrmodeling_tf_utilsrrrrrrrrrrrrr tf_utilsr!r"r#utilsr$r%r&r'r(r)configuration_bertr+ get_loggerrSrrrr:r;r<rOrRrSrPrQr%r&r'r/rgLayerrYrrrrrrr0rLrUr[ryrrrrBERT_START_DOCSTRINGrrrrrrrr)r5r>__all__rWrPrNrds  " !///   SR+   H %5(Z$`"6'+T'(QQ8P u||))P fAH%,,,,AHHLu||))L<0.ell((0.fH++H:L5<<%%L